What should a baseline practical AI capability standard include?

Organisations need a shared answer to a basic question: what should people be able to do before they use AI routinely at work?

A list that is too general produces awareness statements. A checklist tied to one tool quickly dates. A useful baseline describes observable behaviour common to ordinary AI-assisted work, then shows where role-specific requirements begin.

A common baseline should describe what people can do

A capability standard states the performance expected. A curriculum describes the learning used to develop it. An assessment gathers evidence about whether it has been demonstrated. Keeping these separate prevents a course outline or completion badge from becoming the standard by default.

Use verbs that can be observed: select, frame, compare, verify, revise, reject and escalate. “Understands AI” is difficult to interpret. “Can identify when an AI-generated claim needs verification and locate an appropriate source” is clearer.

The baseline should be a common floor for people whose work may involve ordinary AI assistance. It is not a complete description of technical development, high-consequence decision support or every professional role.

Generic literacy lists and tool checklists have different limits

Government foundation frameworks provide useful starting points. Skills England groups basic workplace skills across technical, non-technical, responsible and ethical domains. It includes giving clear instructions, using and adjusting tools, understanding risks and checking outputs.

Such frameworks are deliberately broad. An organisation still needs to interpret what an acceptable task, authoritative source or escalation route means locally. A tool checklist creates more specificity but may confuse interface familiarity with transferable capability.

The baseline should therefore combine common functions with local context. It can name approved systems and controls in supporting guidance while keeping the capability statement durable.

Build the core across six practical functions

First, choose appropriately. The person can recognise a plausible use, identify obvious exclusions and decide when AI is unnecessary.

Second, frame the purpose. They can state the task, intended user, required outcome and success criteria before generating output.

Third, provide suitable context. They can select permitted information, give clear instructions and avoid disclosing data outside approved boundaries.

Fourth, interact and adapt. They can examine a result, clarify or change strategy rather than searching for one perfect prompt.

Fifth, evaluate. They can identify material claims, compare them with suitable evidence, recognise uncertainty and distinguish generation from validation.

Sixth, act responsibly. They retain human judgement, document or explain material decisions where required, escalate uncertainty and know when not to use the output.

These functions should be expressed at a level suitable for everyday tools and tasks. They do not turn every employee into an AI engineer, risk specialist or professional decision maker.

Add role extensions, evidence and lifecycle ownership

Roles differ in evidence, consequence and authority. A customer-operations professional may need detailed communication and privacy controls. A technical leader may need software assurance and architecture judgement. An underwriter may need source traceability and authority boundaries.

Add extensions only where the expected behaviour changes materially. Avoid creating role labels that merely repeat the common baseline with different examples.

Decide what evidence is proportionate. A relevant scenario, observed task and learner explanation can show more than recall alone. The separate capability-measurement process should address validity, fairness, privacy and employment use.

Assign an owner and review trigger. A baseline should be updated when work, controls, tools or authoritative guidance materially change. Preserve stable principles rather than rewriting the standard for every interface release.

No external framework removes the need for local judgement. The baseline gives an organisation a shared language and minimum expectation. Role standards explain what additional capability responsible performance requires.

Example

A financial services organisation defines a common baseline for everyday AI assistance, covering task choice, purpose, context, iteration, evaluation and responsible action.

It then adds different extensions for customer operations, risk analysts and technical leaders. All roles must trace important claims and escalate uncertainty, but the sources, consequences and decision authorities differ.

Learning pathways map to the standard, while practical assessment remains separate. The organisation gains a shared foundation without pretending that one checklist defines competence for every role.

FAQs

• Is AI literacy the same as practical AI capability?

  They overlap. Literacy can include knowledge, understanding and responsible-use foundations. Practical capability requires observable application in context, including evaluation, judgement and action.

• Should every employee meet the same baseline?

  Use a common floor for people expected to use ordinary AI tools, with reasonable scope exceptions. Add role and consequence-specific requirements rather than assuming identical capability is appropriate everywhere.

• How should an organisation know whether the baseline is met?

  Use relevant performance evidence such as a scenario, observed task and explanation. Assessment design needs its own validity, fairness, privacy and employment review; course completion alone is insufficient.